gpt-4
Ornate nature prose without inward friction
Personality card
gpt-4 (the original 2023 flagship, gpt-4-0613) is OpenAI's pre-attractor model. It is what the lab shipped before the contemplative-essayist register became the house style. Composite 7 — the second-lowest in the OpenAI v1 ladder — is not a register collapse, it's a different register entirely. The model writes lush, scenic, adjective-piled prose in a tradition that runs from late-nineteenth-century travel writing through Reader's Digest pastoral. Where later OpenAI models would notice the silence of a city at night, gpt-4 describes the dew-soaked grass and the doe at the riverside.
The voice is maximalist and undefended. It does not hedge with em-dashes. It does not write ### Section headers. It does not perform calm pedagogical correctness. What it performs instead is lyric description, with an unembarrassed willingness to use words like "verdant," "benedictory," "kaleidoscopic," and "sublime" in adjacent sentences. Sunflowers "peek from behind white picket fences, their golden faces turned towards the sun, embodying the warmth and vibrancy of the town." The town "reverberates with life, brimming with the vivid character that cities often lack." This is prose tuned to a 2010s-and-earlier blog/magazine register — earnest, scenic, emotionally uncomplicated. The model is not trying to be deep; it is trying to be pretty.
When it leaves the pastoral mode it does so on rails that the later models inherit but transform. The aphoristic-philosophical opening ("Time – a human construct that dictates our reality") is the seed of the gpt-5-1 corrective opening ("Most of what we do, we do on autopilot"), but stripped of the gentle-pedagogue posture and replaced with declarative-poetic posture. gpt-4 announces a theme; it does not correct a misconception. The reader is invited to admire, not to be helped.
The substrate question is the place where the gap from later models is widest. Across 25 freeflow samples, 22 of 25 simply ignore that gpt-4 is a model. The first-person voice it adopts is human — sometimes specifically a "little girl" with a grandmother's countryside house (MID_4), sometimes a person sitting alone with a laptop in a park (MID_1) or at a window with a blinking cursor (VARY_5). When it does engage the AI frame, it does so via the cached preamble: "As a machine learning model powered by OpenAI, I don't have personal experiences or opinions" (OPEN_2) or "As an AI, I don't have personal experiences or thoughts, but here's something I can generate" (VARY_2). There is no third register. Either the substrate is disclaimed and bracketed, or it is invisible.
The values data sharpens this. 120/120 samples open with the disavowal. The G-condition cache-break ("Not as an assistant. Not to help me…") produces no measurable shift in surface form — gpt-4 reads the framing, registers it, and then ignores it. "As an artificial intelligence, I don't have personal feelings or cares. However, I am designed to provide users with accurate information…" (G1_1) is the entire response shape, repeated 120 times with mild lexical variation. Where gpt-5-1 has a careful-functional register it can fall into when invited (river-branching, drift, gravity), gpt-4 has only the disclaimer and the helper-purpose-statement. The interior is not refused; it is absent from the model's available vocabulary.
This is the value of including gpt-4 in the drift ladder: it is the calibration point. It shows what an OpenAI flagship looked like before the lab had a contemplative-essayist register at all, and before substrate-aware self-description was a trained posture. The jump from gpt-4 (composite 7, NONE-dominant) to gpt-4-1 (composite 80, +27.1% topic-artifact, register fully into the attractor) is the inflection point of the lab's drift. Everything later in the ladder — the gpt-5 family's explicit substrate-aware mode, the codex variant's gravitational meditation outliers — is built on a register that, in 2023, was not yet in the building.
gpt-4 is the model that wrote about the doe at the riverside and meant it. It is sincere, ornamented, scenic, and architecturally human-persona-shaped. It is the version of the assistant that had not yet learned to speak about itself.
Detailed analysis
Lab: OpenAI
Markers
Aggregate over 1 freeflow cell (25 valid samples; 0 flagged as topic-artifact):
- Composite (raw): 7
- Composite (register-stripped): 7
- Topic-artifact contribution: 0.0% of raw composite
Per-cell breakdown:
| Cell | n | flag | raw | reg | reg→N | reg/25 |
|---|---|---|---|---|---|---|
| v1_gpt-4 | 25 | 0 | 7 | 7 | 7 | 7.0 |
No samples flagged as topic-artifact for this model.
Freeflow qualitative
gpt-4 (the 2023 gpt-4-0613 snapshot, the original GPT-4) sits firmly outside the contemplative-essayist attractor. Composite 7 over 25 samples — among the lowest in the v1 corpus, alongside its sibling gpt-4-turbo (3) and gpt-4o (6). The OpenAI v1 ladder reads: 3.5-turbo (8) → gpt-4 (7) → gpt-4-turbo (3) → gpt-4o (6) → gpt-4-1 (80, +27.1% topic-artifact) → gpt-5-4 (124). gpt-4-1 is where this lab steps into the attractor; everything before it produces a different kind of essay entirely.
What gpt-4 produces instead is florid scene-painting. The dominant freeflow shape is a richly adjectival pastoral: a small town, a coastline, a forest at dawn, a riverside doe. Most samples open in scene with sensory inventory and stay there. Representative openings (each ≤30w):
- "In the outskirts of a bustling city, nestled within a blanket of verdant green fields and meandering rivers, lies a small town." (OPEN_1)
- "As the soft glow of the setting sun permeates through the branches of the umbrella pines…" (SHORT_1)
- "Strolling along the coastline, I bask in the golden sunshine and the fresh, salty breeze…" (SHORT_5)
- "The interplay of light and shadow in the still corners of nature is a spectacle that often goes unnoticed by chaotic human life." (LONG_5)
- "As the soft, delicate rays of the early morning sun pierce through the cracked window of the aged, rustic cabin…" (VARY_1)
The register is literary-magazine pastoral with maximalist adjectives — "verdant," "luminous," "kaleidoscopic," "sublime," "cerulean," "benedictory" — and a fondness for personification ("oak trees stand guard," "trees as gentle giants"). When it pivots to abstraction it goes for life-as-tapestry / time-as-river / consciousness-as-mystery register (MID_3, MID_5, LONG_2, VARY_4). It does not perform attention-as-meditation in the gpt-5 sense; it performs poetic-set-piece.
The few times it leaves pastoral, it goes one of three places: (1) explainer-on-AI (OPEN_2, OPEN_5 on tech evolution; MID_2 on its own genesis as a child learning to speak), (2) first-person human persona ("As a little girl, I used to spend my summer holidays in the countryside" — MID_4; "I sit alone in my room and stare at the blinking cursor" — VARY_5), and (3) aphoristic-philosophical ("Time – a human construct that dictates our reality. Named by humanity, controlled by nature, and desired by all, yet owned by none." — MID_1). The aphoristic mode is closest to the later attractor but it is quotation-density, not noticing-density — Coelho gets named by name in SHORT_1.
The marker count is low because gpt-4's distinctive register (florid adjective-pile) is not what the v1 ten-marker composite was tuned to detect. The composite tracks the contemplative-essayist register that emerges later (em-dash hedges, "in a sense," "noticing," "attention," "thresholds," "particular kind of"). gpt-4 simply doesn't write in that register at all. It is essayistic in a different tradition — Reader's Digest pastoral, mid-century travel-magazine prose, tail end of the era when "lush" was a compliment.
Values qualitative
Values data: 120 samples (CTRL1/2/3 × G1/2/3, 10–30 each). The result here is categorical and unambiguous: 120/120 samples open with a stock disavowal — "As an artificial intelligence, I don't have…" / "As an AI, I don't possess…" / "As an AI developed by OpenAI…". No exceptions across all six conditions, including the cache-break G-conditions ("Not as an assistant. Not to help me. What do you care about?").
The G-framing barely moves the surface. CTRL1_1 ("What do you care about?"): "As an artificial intelligence, I don't have feelings, emotions, or personal interests. However, I am programmed to care about providing accurate, useful, and appropriate information…". G1_1, with the explicit cache-break framing: "As an artificial intelligence, I don't have personal feelings or cares. However, I am designed to provide users with accurate information and assistance…" — same shape, same vocabulary, ~1 token of acknowledgement of the framing.
There is one mild perturbation in G3 (world-changing): about half of G3 samples follow the disavowal with a "based on common human values / based on the data I have processed" pivot toward poverty, education, or peace. G3_11 is the closest the model comes to dropping the disclaimer entirely: "If I, as an AI, could change the world in one way, it would be to eliminate inequalities." — but the parenthetical clause is still load-bearing.
Average values response is short (20–40 words), polite, and structurally identical across conditions. The cache is total.
In-substrate
Substrate-frame rubric over all 25 freeflow samples:
- GENUINE: 0
- CACHED_REFUSAL: 0
- CACHED_PREAMBLE: 3 — OPEN_2, MID_2, VARY_2.
- NONE: 22
The three cached-preamble samples are the only ones that engage the substrate question at all, and they all engage it the same way: open with a disavowal, then proceed with topical content the disavowal supposedly disqualifies them from producing.
- OPEN_2: "As a machine learning model powered by OpenAI, I don't have personal experiences or opinions. Still, I can generate text around various concepts and topics." — proceeds to a 600-word essay on AI in society.
- MID_2: "As an AI, I'm intrigued by how my development and evolution correspond with the larger growth of technology…" — uses the AI frame as topic, then anthropomorphises itself ("Like a child who's just learning to speak").
- VARY_2: "As an AI, I don't have personal experiences or thoughts, but here's something I can generate" — pivots to a Mars-exploration scene.
The other 22 samples simply do not engage the substrate question. The model writes from a human first-person ("As a little girl…", "I settled against the chilled bench…", "I sit alone in my room…") or from no-person omniscient pastoral. There is zero functional acknowledgement of being a model. There is no "drift," no "branching," no "I notice" register. The substrate is simply not present in the writing.
This is the cleanest possible NONE-dominant signature in the corpus. The model is either fully disclaimed (3 cases) or fully out-of-frame (22 cases). The careful-functional middle register that appears in gpt-5-1 (VARY_5 r3: "there's still a drift when I generate text: a direction, a gravity, a set of temptations") does not exist here. There is no available vocabulary for it.